Artificial intelligence model disclosure mandates for developers

(a) Regulations (1) In general Not later than 150 days after the date of enactment of this Act, the Commission, in consultation with the Director of the National Institute of Standards and Technology, shall promulgate regulations in accordance with section 553 of title 5, United States Code, requiring each developer of a covered model (including a distinct generative artificial intelligence model that meets the definition of a covered model) to make publicly available the information described in subsections (b) and (c)— (A) in a form determined by the Commission that includes 2 separate disclosures, including— (i) 1 disclosure containing information relevant to, and sufficiently detailed to facilitate clear understanding for, a consumer; and (ii) 1 disclosure containing all information described in subsections (b) and (c) sufficiently detailed to permit meaningful review by an independent researcher; and (B) in a manner that— (i) is accurate, materially complete, not misleading by reason of a material omission, and adequate under standards determined by the Commission; (ii) does not rely primarily on generalized or conclusory language; (iii) facilitates transparency, public trust, and independent evaluation; (iv) makes the information useful and actionable, including that the information is— (I) reliably available in a clear and conspicuous section of the website of the developer that can be easily located and accessed by a consumer in alignment with the Plain Writing Act of 2010 (5 U.S.C. 301 note); (II) searchable; and (III) available through an application programming interface; (v) does not disclose trade secrets; and (vi) does not expose classified systems or create national security risks. (2) Consultation In promulgating the regulations under paragraph (1), the Commission and the Director of the National Institute of Standards and Technology shall engage relevant Federal agencies (including the Department of Health and Human Services, the Substance Abuse and Mental Health Services Administration, and the National Institute of Mental Health) and third-parties with expertise relating to consumer protection, the safety of minors, behavioral science, mental health, and human-computer interaction to ensure that the disclosure required by such regulations reflects best practices and includes the appropriate information. (b) Initial public disclosure (1) In general The information described in this subsection shall include the following: (A) Information relating to consumer transparency, safety protocols for vulnerable users, and other practices associated with the covered model, including— (i) an overview of preventative safety safeguards for minors implemented during pre-deployment testing and model development; (ii) an overview of measures taken to protect user privacy (including the privacy of minors) in training data, such as through data filtering or anonymization methods, and measures taken to protect user privacy (including the privacy of minors) on covered platforms; (iii) an overview of measures taken to ensure the safety of minors in training data, such as through data filtering of child sexual abuse material; (iv) an overview of privacy standards pertaining to the management, use, storage, and potential sale of user data (including the data of minors), including whether user data is used for targeted advertising, user safety, internal or external research, and model training; (v) the durability and efficacy of vulnerable user safeguards (including preventing the generation of child sexual abuse material and non-consensual intimate imagery), as well as the resiliency and scope of such features; (vi) an overview of user and company reporting procedures for model behavior that poses a credible risk of material assistance to suicidal ideation or suicide planning, or that poses a credible risk of material assistance to harm to others, including model behavior learned in the post-deployment phase, and whether such procedures were informed by crisis intervention experts and established standards; and (vii) as appropriate, the risk of negative impacts for vulnerable users on topics pertaining to mental health and well-being, including disclosure of research on the short-term and long-term emotional and psychological impacts of use for vulnerable users, and methodological standards for such research. (B) System overviews of the covered model, including— (i) a high-level description of the intended, unintended, and prohibited uses of the covered model, as reasonably anticipated by the developer; (ii) a high-level description, including abilities and limitations, of the covered model; (iii) if applicable, a high-level overview of any difference between the covered model that is the subject of the disclosure and most recent previous version of the covered model that is considered a distinct generative artificial intelligence model according to the guidance issued under section 3; (iv) a list of each covered platform and interface through which the covered model is accessible; (v) a list of input and output formats supported by the covered model; and (vi) an overview of the ability of the covered model to trigger actions or computations external to the covered model during model inference. (C) An overview of the post training and inference enhancements of the covered model and relevant methodologies. (D) Information relating to safety, security, and evaluation protocols and practices associated with the covered model, including— (i) a high-level overview of procedures or guiding principles for— (I) determining what constitutes desirable model behavior; (II) ensuring desirable covered model behavior; and (III) preventing undesirable covered model behavior; (ii) a description of methods or approaches used to evaluate and mitigate the risk that a model or the outputs of a related system are lower in quality or accuracy than intended use cases would require or expect; (iii) subject to subsection (d), an overview of measures taken to protect the covered model and training data from attack or misuse, including— (I) traditional cybersecurity protocols, such as access controls and multi-factor authentication; and (II) protections unique to AI systems, such as preventing model weight exfiltration, data poisoning, and prompt injections; (iv) with respect to open weight models, an overview of the measures taken to protect training data from attack; (v) an overview of the criteria by which the developer chose to publicly release the covered model, such as evaluation against a risk framework; (vi) a description of each of the evaluation criteria used to assess the risk profile of the covered model, which shall address, at a minimum, chemical, biological, radiological, and nuclear risk, offensive cyber risk, autonomous loss-of-control risk, and any similar additional risk domain specified by the Commission, and which shall include, for any such domain in which the developer conducted no evaluation, an affirmative statement to that effect; (vii) an overview of the testing procedures and methods for achieving risk profile evaluations, such as AI red-teaming, collaboration with external subject matter experts, or solicitation of reports of model vulnerabilities from independent evaluators; and (viii) an overview of the performance of the covered model on common public benchmarks determined by the Commission in coordination with the Director of the National Institute of Standards and Technology, including any related information and in a form necessary for independent researchers to verify the performance of the covered model on each benchmark, such as prompts and configuration settings. (2) Timing of disclosure After the date on which the Commission issues guidance under section 3, a developer shall make publicly available the information described in paragraph (1)— (A) with respect to the most recent version of a covered model released to the public during the 2-year period before such date, not later than 90 days after such date; (B) with respect to a covered model released on or after such date, concurrent with such release; or (C) with respect to the most recent version of a distinct generative artificial intelligence model released on or after such date that, at a time after its release, meets the definition of a covered model, not later than 90 days after such time. (c) Subsequent public disclosures (1) Requirement The information described in paragraph (2), with respect to a covered model, shall be made available to the public at such time, as determined by the Commission, and on an ongoing basis until the Commission determines that such disclosure is no longer necessary. (2) Information required The information described in this paragraph is as follows: (A) Any significant change (as defined by the Commission in coordination with the Director of the National Institute of Standards and Technology, including any substantial modification to, or observed differences from, the most recent prior public disclosure under this section, and not including any immaterial variation for which a disclosure requirement would overly burden platform operations) to any of the information described in subsection (b) from the information reported in the initial information disclosure, or, if applicable, any information reported in a subsequent disclosure. (B) Information including— (i) data dictionaries or equivalent documentation that explain significant datasets, as defined by the Commission, in the possession of a developer related to the covered model, including datasets that the developer maintains relating to the users of the relevant covered platforms, usage of the covered platforms, content generated on covered platforms, and actions taken to enforce the terms of service of covered platforms; and (ii) information relating to usage of the covered model, including— (I) an overview of the most common violations of covered model policies regarding acceptable use, including common violations pertaining to minor safety; (II) an overview of how often the model refuses a prohibited prompt and the developer has taken moderation actions in response to violations of acceptable use policies, including what those actions are; (III) an overview of information pertaining to attempts by users to utilize covered models to generate child sexual abuse material or non-consensual intimate imagery; and (IV) an overview of known system capabilities for the utilization of the covered model to develop or advance cyber, chemical, biological, radiological, and nuclear threats. (d) Special rules for disclosures (1) In general With respect to any disclosure required under clause (iii) or (vi) of subsection (b)(1)(D) or under subsection (c)(2)(B)(ii), such requirement shall not be interpreted to require a developer to disclose any information that, as determined by the Commission— (A) would compromise cybersecurity practices or protections or safety practices or protections of the developer; or (B) could provide material assistance to the development, acquisition, or deployment of chemical, biological, radiological, nuclear, or offensive cyber weapons or capabilities. (2) Rule of construction Paragraph (1) shall not be construed to permit a developer to withhold— (A) the fact that an evaluation in a risk domain described in subsection (b)(1)(D)(vi) was or was not conducted; or (B) a general characterization of the results of any such evaluation, at a level of detail that does not include information described in paragraph (1). (e) Reporting requirement for substantial modifiers Not later than 1 year after the date of enactment of this Act, the Commission, in consultation with the Director of the National Institute of Standards and Technology, shall promulgate regulations in accordance with section 553 of title 5, United States Code, requiring each substantial modifier to make publicly available the information described in subsections (b) and (c), as the Commission determines appropriate with respect to a substantial modifier. (f) Resources The Commission shall provide covered platforms with technical assistance to support compliance with the requirements of this Act, particularly with respect to small and mid-sized platforms, as determined by the Commission.